NeurIPS 2025 Papers — Page 7
Conference on Neural Information Processing Systems · 5275 papers
Breaking the Performance Ceiling in Reinforcement Learning requires Inference Strategies
Felix Chalumeau (InstaDeep), Arnu Pretorius (Stellenbosch University)
Reinforcement Learning
🎯 What it does: This study investigates search strategies during the reasoning phase of reinforcement learning, significantly improving performance in multi-agent tasks by using various reasoning strategies such as random sampling, tree search, online fine-tuning, and COMPASS.
Breakthrough Sensor-Limited Single View: Towards Implicit Temporal Dynamics for Time Series Domain Adaptation
Mingyang Liu (Harbin Institute of Technology), Liqiang Nie (Harbin Institute of Technology)
Domain AdaptationConvolutional Neural NetworkTime Series
🎯 What it does: The EDEN framework is proposed, which performs unsupervised domain adaptation on time series using multiple explicit domains (multi-scale, multi-subspace, multi-paragraph) and integrates three specialized modules to achieve more robust domain-invariant representations.
BridgePure: Limited Protection Leakage Can Break Black-Box Data Protection
Yihan Wang (University of Waterloo), Yaoliang Yu (University of Waterloo)
ClassificationRestorationSafty and PrivacyDiffusion modelImage
🎯 What it does: Proposes the BridgePure method, which utilizes a small amount of protective leakage for reverse mapping of black-box data protection to restore data availability.
BridgeVLA: Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models
Peiyan Li (Institute of Automation, Chinese Academy of Sciences), Tieniu Tan (Institute of Automation, Chinese Academy of Sciences)
Robotic IntelligenceConvolutional Neural NetworkTransformerReinforcement LearningVision Language ModelVision-Language-Action ModelImagePoint Cloud
🎯 What it does: This paper presents BridgeVLA, a 3D vision-language-action (VLA) framework based on a pre-trained vision-language model (VLM) that enables efficient 3D robotic manipulation learning by aligning inputs and outputs in a multi-view 2D heatmap space.
Bridging Arbitrary and Tree Metrics via Differentiable Gromov Hyperbolicity
Pierre Houédry, Titouan Vayer (INRIA)
OptimizationGraph
🎯 What it does: A differentiable optimization framework DELTAZERO is proposed to minimize the distortion from any metric space to tree metrics, outputting an approximate tree metric.
Bridging Brains and Concepts: Interpretable Visual Decoding from fMRI with Semantic Bottlenecks
Sara Cammarota (University of Rome Tor Vergata), Nicola Toschi (University of Rome Tor Vergata)
GenerationExplainability and InterpretabilityLarge Language ModelDiffusion modelImageMagnetic Resonance Imaging
🎯 What it does: An interpretable visual brain decoding framework was constructed, mapping fMRI signals to a 214-dimensional binary concept space, then to CLIP embeddings, and finally generating images using BrainDiffuser.
Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts
Chaitanya Kapoor (University of California), Meenakshi Khosla (University of California)
Domain AdaptationRepresentation LearningConvolutional Neural NetworkTransformerImage
🎯 What it does: This paper systematically evaluates the representation similarity (i.e., convergent learning) of different deep learning models under training processes, network architectures, and distribution shifts, and verifies its patterns through large-scale experiments.
Bridging Equivariant GNNs and Spherical CNNs for Structured Physical Domains
Colin Kohler (Northeastern University), Robin Walters (Northeastern University)
Super ResolutionConvolutional Neural NetworkGraph Neural NetworkReinforcement LearningContrastive LearningPoint CloudMeshPhysics Related
🎯 What it does: This paper proposes G2Sphere, a framework that utilizes SO(3) equivariant graph networks and spherical CNNs to perform encoding and decoding in Fourier space, directly mapping 3D geometry to continuous, high-frequency spherical signals, achieving zero-step super-resolution and efficient inference.
Bridging Expressivity and Scalability with Adaptive Unitary SSMs
Arjun Karuvally (Salk Institute for Biological Studies), Hava T Siegelmann
Time SeriesSequentialOrdinary Differential Equation
🎯 What it does: An Adaptive Unitary State Space Model (AUSSM) is proposed, achieving a balance between expressiveness and scalability for long sequences through input-dependent skew-symmetric recursion.
Bridging Human and LLM Judgments: Understanding and Narrowing the Gap
Felipe Maia Polo (University of Michigan), Yuekai Sun (University of Michigan)
TransformerLarge Language ModelTextMultimodalityBenchmarkChain-of-Thought
🎯 What it does: A Bridge framework has been established to unify the modeling of human and LLM scoring, and through this model, achieve calibration of LLM scoring and quantitative analysis of systematic differences.
Bridging Scales: Spectral Theory Reveals How Local Connectivity Rules Sculpt Global Neural Dynamics in Spatially Extended Networks
Yuhan Huang (Peking University), Guozhang Chen (Peking University)
GraphPhysics RelatedStochastic Differential Equation
🎯 What it does: A two-dimensional spatially extended E/I network model was constructed, and various macroscopic dynamical states such as from no synchronization to synchronization, waves, spots, spikes, and chaos were quantitatively predicted through spectral theory (random matrix theory + Fourier transform);
Bridging Sign and Spoken Languages: Pseudo Gloss Generation for Sign Language Translation
Jianyuan Guo (City University of Hong Kong), Trevor Cohn (Google)
TransformerLarge Language ModelVideoText
🎯 What it does: Utilize large language models (LLM) to generate pseudo glosses and construct a sign language translation (SLT) framework without gloss annotations.
Bridging Symmetry and Robustness: On the Role of Equivariance in Enhancing Adversarial Robustness
Longwei Wang (University of South Dakota), Yang Zhou (Auburn University)
Adversarial AttackConvolutional Neural NetworkImage
🎯 What it does: Embedded rotation and scale equivariant convolutional layers into traditional CNNs, constructed and evaluated various symmetry-enhanced architectures to improve adversarial robustness without relying on adversarial training.
Bridging the Gap Between Cross-Domain Theory and Practical Application: A Case Study on Molecular Dissolution
Sihan Wang (University of Science and Technology of China), Yang Wang (University of Science and Technology of China)
OptimizationDrug DiscoveryGraph Neural NetworkSupervised Fine-TuningMultimodalityGraph
🎯 What it does: This paper proposes Core Subset Iterative Extraction (CSIE) and Asymmetric Molecular Interaction Graph Neural Network (AMGNN) to transfer large-scale theoretical computational data to a limited experimental solvation free energy dataset.
Bridging the gap to real-world language-grounded visual concept learning
Whie Jung (Korea Advanced Institute of Science and Technology), Seunghoon Hong (Korea Advanced Institute of Science and Technology)
GenerationData SynthesisTransformerVision Language ModelDiffusion modelImage
🎯 What it does: A scalable visual concept learning framework is proposed, capable of adaptively discovering and aligning diverse language-specified concept axes in real-world scenarios, and achieving concept decoupling through combinatorial anchoring.
Bridging Theory and Practice in Link Representation with Graph Neural Networks
Veronica Lachi (Fondazione Bruno Kessler), Manfred Jaeger (University of Aalborg)
Representation LearningGraph Neural NetworkContrastive LearningGraphBenchmark
🎯 What it does: This paper proposes a unified kϕ‑kρ‑m framework to theoretically analyze the expressive power of graph neural networks in link representation and designs a synthetic benchmark LR‑EXP specifically for evaluating link expressive power. It then uses graph symmetry metrics to evaluate real link prediction datasets, demonstrating that models with higher expressive power perform better on graphs with higher symmetry.
Bridging Time and Linguistics: LLMs as Time Series Analyzer through Symbolization and Segmentation
Jianyang Qin (Harbin Institute of Technology), Qing Liao (Harbin Institute of Technology)
ClassificationAnomaly DetectionRecurrent Neural NetworkTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTime SeriesSequential
🎯 What it does: By performing spectral symbolization and segment position encoding in the frequency domain, the time series is converted into a text format that can be input into LLMs, and pre-trained LLMs are utilized for tasks such as time series prediction and classification.
Bringing SAM to new heights: leveraging elevation data for tree crown segmentation from drone imagery
Mélisande Teng (Mila - Quebec Artificial Intelligence Institute), Hugo Larochelle (Mila - Quebec Artificial Intelligence Institute)
Object DetectionSegmentationConvolutional Neural NetworkTransformerPrompt EngineeringImage
🎯 What it does: This paper studies the combination of the Segment Anything Model (SAM) and Digital Surface Model (DSM) for tree crown instance segmentation in high-resolution drone imagery, evaluating it in three ecological contexts: northern coniferous forests, temperate forests, and tropical rainforests.
Broken Tokens? Your Language Model can Secretly Handle Non-Canonical Tokenizations
Brian Siyuan Zheng (University of Washington), Noah A. Smith (Allen Institute for AI)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: This study explores the robustness of language models when faced with non-canonical tokenization methods that were not seen during training, demonstrating that instruction-tuned models maintain a high level of performance; it also evaluates whether changing tokenization strategies during inference can enhance model performance.
BTL-UI: Blink-Think-Link Reasoning Model for GUI Agent
Shaojie Zhang (Xiaomi Inc), Jian Luan (Xiaomi Inc)
Robotic IntelligenceLarge Language ModelReinforcement LearningVision Language ModelMultimodality
🎯 What it does: Designed and implemented the Blink-Think-Link (BTL) framework and trained the BTL-UI GUI agent, significantly enhancing the perception, reasoning, and execution performance of GUI interactions.
Buffer layers for Test-Time Adaptation
Hyeongyu Kim (Yonsei University), Dosik Hwang (Korea Institute of Science and Technology)
Domain AdaptationConvolutional Neural NetworkImage
🎯 What it does: A Buffer layer is proposed, which is a lightweight, pluggable auxiliary module for source-agnostic online testing adaptation, keeping the original backbone network unchanged;
Building 3D Representations and Generating Motions From a Single Image via Video-Generation
Weiming Zhi (University of Sydney), Matthew Johnson-Roberson (Carnegie Mellon University)
GenerationData SynthesisRobotic IntelligenceContrastive LearningImageVideo
🎯 What it does: Construct a complete 3D environment representation from a single RGB image and achieve collision-avoidance motion trajectory generation based on this.
BundleFlow: Deep Menus for Combinatorial Auctions by Diffusion-Based Optimization
Tonghan Wang (Harvard University), David C. Parkes (Harvard University)
OptimizationFlow-based ModelTabularBenchmarkOrdinary Differential Equation
🎯 What it does: This paper proposes a menu-based combinatorial auction design framework based on Continuous Regularized Flows (ODE) - BUNDLEFLOW, aimed at achieving revenue-optimal auctions that are DSIC in a single-buyer setting.
C-LoRA: Contextual Low-Rank Adaptation for Uncertainty Estimation in Large Language Models
Amir Hossein Rahmati (Texas A&M University), Xiaoning Qian (Brookhaven National Laboratory)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: Proposes Contextual Low-Rank Adaptation (C-LoRA), which incorporates a data-dependent lightweight context module into the LoRA fine-tuning of LLMs, achieving scalable Bayesian uncertainty estimation.
C-NAV: Towards Self-Evolving Continual Object Navigation in Open World
MingMing Yu, Jing Liu
Knowledge DistillationRobotic IntelligenceRecurrent Neural NetworkTransformerReinforcement LearningMultimodalityPoint CloudBenchmark
🎯 What it does: A benchmark for goal navigation in continual learning is proposed, and the C-Nav framework is designed to learn new target categories while avoiding forgetting old ones.
C-SafeGen: Certified Safe LLM Generation with Claim-Based Streaming Guardrails
Mintong Kang (University of Illinois at Urbana-Champaign), Bo Li (University of Illinois at Urbana-Champaign)
GenerationSafty and PrivacyComputational EfficiencyTransformerLarge Language ModelText
🎯 What it does: The C-SafeGen framework is proposed, achieving a provable upper bound on the safety risks of LLM-generated text and configuration certification; based on this, the Claim-Based Stream Decoding (CSD) algorithm is designed to dynamically monitor and backtrack potential unsafe 'claims' during the generation process, utilizing KV caching to enhance efficiency.
C$^2$Prompt: Class-aware Client Knowledge Interaction for Federated Continual Learning
Kunlun Xu (Peking University), Jiahuan Zhou (Peking University)
Federated LearningKnowledge DistillationTransformerPrompt EngineeringImage
🎯 What it does: This paper proposes C Prompt, a class-aware client knowledge interaction method for federated continual learning, which achieves knowledge consistency across clients through local class distribution compensation and class-aware prompt aggregation.
C3PO: Optimized Large Language Model Cascades with Probabilistic Cost Constraints for Reasoning
Antonios Valkanas (McGill University), Mark Coates (McGill University)
TransformerLarge Language ModelTextBenchmark
🎯 What it does: A self-supervised LLM cascading reasoning framework C3PO is designed, which uses a confidence threshold to decide whether to exit early and controls reasoning costs through conformal prediction.
CAD-Coder: Text-to-CAD Generation with Chain-of-Thought and Geometric Reward
Yandong Guan (Beihang University), Qian Yu (Hong Kong University)
GenerationReinforcement Learning from Human FeedbackTransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningTextChain-of-Thought
🎯 What it does: The CAD-Coder framework is proposed, which generates 3D CAD models from natural language using CadQuery scripts.
CADGrasp: Learning Contact and Collision Aware General Dexterous Grasping in Cluttered Scenes
Jiyao Zhang (Peking University), Hao Dong (Peking University)
OptimizationRobotic IntelligenceDiffusion modelPoint Cloud
🎯 What it does: A two-stage framework called CADGrasp is proposed, using sparse interactive bisectors (IBS) as a scene-independent intermediate representation for contact and collision perception. It first predicts IBS and then obtains stable, collision-free flexible grasp poses through energy optimization.
CADMorph: Geometry‑Driven Parametric CAD Editing via a Plan–Generate–Verify Loop
Weijian Ma (Fudan University), Jiang Bian (Microsoft Research)
Large Language ModelDiffusion modelMesh
🎯 What it does: This paper presents CADMorph, a geometry-driven parametric CAD editing framework based on a plan-generate-validate loop.
CAGE: Continuity-Aware edGE Network Unlocks Robust Floorplan Reconstruction
Yiyi Liu (Wuhan University of Technology), Biao Xiong (Wuhan University of Technology)
TransformerPoint Cloud
🎯 What it does: A continuous perceptual edge (CAGE) network is proposed, which directly reconstructs vectorized indoor floor plans using 2D density maps projected from point clouds.
Calibrating Translation Decoding with Quality Estimation on LLMs
Di Wu (University of Amsterdam), Christof Monz (University of Amsterdam)
TransformerLarge Language ModelSupervised Fine-TuningText
🎯 What it does: The paper achieves the calibration of the translation process of large language models by directly optimizing the likelihood of the translation model and the Pearson correlation with translation quality during the training phase, significantly improving translation quality and quality estimation capabilities.
CaliGCL: Calibrated Graph Contrastive Learning via Partitioned Similarity and Consistency Discrimination
Yuena Lin (Beijing University of Technology), Gengyu Lyu (Beijing University of Technology)
ClassificationRepresentation LearningGraph Neural NetworkContrastive LearningGraph
🎯 What it does: This paper proposes CaliGCL, which addresses the estimation bias of similarity and the semantic shift bias in graph contrastive learning. It designs an exponential partition similarity and a semantic consistency discriminator to correct contrastive supervision and improve the quality of unsupervised graph representation learning.
CALM-PDE: Continuous and Adaptive Convolutions for Latent Space Modeling of Time-dependent PDEs
Jan Hagnberger (University of Stuttgart), Mathias Niepert (University of Stuttgart)
Convolutional Neural NetworkTransformerTime SeriesPhysics Related
🎯 What it does: This paper proposes a compressed latent space model CALM-PDE based on continuous adaptive convolution for efficiently solving arbitrary discrete time-varying partial differential equations.
CALM: Culturally Self-Aware Language Models
Lingzhi Shen (University of Southampton), Shoaib Jameel (University of Southampton)
TransformerLarge Language ModelMixture of ExpertsContrastive LearningTextMultimodality
🎯 What it does: Proposes the CALM framework, enabling large language models to self-perceive during inference and integrate cultural knowledge for cross-cultural adaptation.
CAM: A Constructivist View of Agentic Memory for LLM-Based Reading Comprehension
Rui Li (Renmin University of China), Ruiming Tang (Huawei)
GenerationRetrievalTransformerLarge Language ModelAgentic AITextRetrieval-Augmented Generation
🎯 What it does: Proposed and implemented a Constructive Agent Memory System (CAM) based on constructivist theory for long text reading comprehension.
CamEdit: Continuous Camera Parameter Control for Photorealistic Image Editing
Xinran Qin (Shenzhen Campus of Sun Yat-sen University), Xiaochun Cao
GenerationData SynthesisTransformerPrompt EngineeringDiffusion modelImage
🎯 What it does: Proposes the CamEdit framework, which utilizes continuous text prompts to perform high-fidelity, optically realistic editing of camera parameters (aperture, focus plane, shutter speed) for real images.
Cameras as Relative Positional Encoding
Ruilong Li (University of California Berkeley), Angjoo Kanazawa (University of California Berkeley)
Data SynthesisDepth EstimationTransformerImage
🎯 What it does: Proposes using the camera as a relative position encoding, introducing a new Projective Positional Encoding (PRoPE) and applying it to multi-view Transformers to enhance geometric reasoning capabilities.
CAMILA: Context-Aware Masking for Image Editing with Language Alignment
Hyunseung Kim (Samsung Semiconductor), Joon Hee Choi (Samsung Semiconductor)
GenerationData SynthesisTransformerLarge Language ModelDiffusion modelImageMultimodality
🎯 What it does: A context-aware image editing method called CAMILA is proposed, which utilizes a multimodal large language model to generate [MASK]/[NEG] tokens. It generates editing masks through a Token Broadcaster and Token Decoder, and achieves precise editing of executable instructions while ignoring non-executable ones using a diffusion model.
CAML: Collaborative Auxiliary Modality Learning for Multi-Agent Systems
Rui Liu (University of Maryland), Ming Lin (University of Maryland)
SegmentationAnomaly DetectionAutonomous DrivingKnowledge DistillationRobotic IntelligenceMultimodality
🎯 What it does: This paper proposes a framework called Collaborative Auxiliary Modality Learning (CAML), which utilizes multiple agents to share multimodal data during the training phase and achieves inference using only a small number of modalities through knowledge distillation.
CAMO: Convergence-Aware Multi-Fidelity Bayesian Optimization
WEI W. XING, Akeel Shah
OptimizationTabularStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: A convergence-aware multi-fidelity Bayesian optimization framework CAMO based on Linear Fidelity Differential Equations (LiFiDE) is proposed, which can explicitly model the convergence process of the objective function at different fidelities and achieve efficient multi-fidelity modeling through adaptive kernel functions.
CamSAM2: Segment Anything Accurately in Camouflaged Videos
Yuli Zhou (ETH Zurich), Guolei Sun (Nankai University)
SegmentationVideoMultimodality
🎯 What it does: This paper improves the performance of SAM2 in the task of video camouflage object segmentation by introducing learnable de-camouflage tokens, implicit and explicit object perception fusion modules, and an object prototype generation module, while keeping the original parameters of SAM2 unchanged.
Can Agent Fix Agent Issues?
Alfin Wijaya Rahardja (Fudan University), Yiling Lou (University of Illinois Urbana-Champaign)
TransformerLarge Language ModelAgentic AITextBenchmark
🎯 What it does: This paper first conducts a manual analysis of 201 GitHub issues from 18 real LLM agent systems, constructing a classification system for agent system issues that includes 6 categories (compatibility, tools, memory, workflow, LLM operations, practicality) and 20 subcategories. Based on 50 reproducible issues, the AGENTISSUE-BENCH benchmark is manually built, and then automated repair experiments are conducted on this benchmark using three mainstream software engineering (SE) agents (SWE-agent, AutoCodeRover, Agentless) combined with GPT-4o and Claude-3.5-Sonnet.
Can Class-Priors Help Single-Positive Multi-Label Learning?
Biao Liu (Southeast University), Xin Geng (Southeast University)
ClassificationImage
🎯 What it does: This paper proposes a framework named CRISP for Single Positive Sample Multi-Label Learning (SPMLL), which first estimates the prior probabilities of each category and then uses the estimated priors to construct an unbiased risk estimator to train the multi-label classifier.
Can Dependencies Induced by LLM-Agent Workflows Be Trusted?
Yu Yao (Peking University), Tongliang Liu (University of Sydney)
Large Language ModelAgentic AITextMultimodality
🎯 What it does: Proposes the SEQCV framework to address the issue of agent misconfiguration in multi-LLM agent systems;
Can Diffusion Models Disentangle? A Theoretical Perspective
Liming Wang (Massachusetts Institute of Technology), James R. Glass (Massachusetts Institute of Technology)
ClassificationGenerationRepresentation LearningDiffusion modelScore-based ModelImageAudio
🎯 What it does: A distinguishable representation learning framework based on diffusion models is proposed, providing a distinguishability theory under weak supervision and validating it on tasks such as Gaussian mixture, image colorization, denoising, and speech emotion classification.
Can DPO Learn Diverse Human Values? A Theoretical Scaling Law
Shawn Im (University of Wisconsin Madison), Sharon Li (University of Wisconsin Madison)
TransformerLarge Language ModelReinforcement LearningText
🎯 What it does: This paper explores the generalization performance of Direct Preference Optimization (DPO) under diverse human values and proposes a dynamic analysis framework based on reward margins for theoretical derivation and experimental validation.
Can Knowledge-Graph-based Retrieval Augmented Generation Really Retrieve What You Need?
Junchi Yu (University of Oxford), Dongzhan Zhou (Shanghai Artificial Intelligence Laboratory)
RetrievalGraph Neural NetworkLarge Language ModelTextGraphBenchmarkRetrieval-Augmented Generation
🎯 What it does: A new framework called GraphFlow is proposed, which achieves multi-step retrieval of text-rich knowledge graphs through the joint optimization of retrieval strategies and flow estimators.
Can Large Language Models Master Complex Card Games?
Wei Wang (Nankai University), Jie Tang (Tsinghua University)
TransformerLarge Language ModelSupervised Fine-TuningReinforcement LearningText
🎯 What it does: By supervising the fine-tuning of LLM on high-quality game data, this study explores and verifies whether LLM can master eight complex card games, and examines its ability to master multiple games simultaneously as well as the impact on general capabilities.
Can LLMs Reason Over Non-Text Modalities in a Training-Free Manner? A Case Study with In-Context Representation Learning
Tianle Zhang (Nanyang Technological University), Alvin Chan (Nanyang Technological University)
Representation LearningDrug DiscoveryTransformerLarge Language ModelPrompt EngineeringMultimodalityBiomedical Data
🎯 What it does: This paper proposes an unsupervised training framework for In-Context Representation Learning (ICRL), which can directly inject features from non-text modalities (such as molecular representations) into large language models (LLMs) during inference, enabling multimodal reasoning.
Can MLLMs Absorb Math Reasoning Abilities from LLMs as Free Lunch?
Yijie Hu (Duke Kunshan University), Qiufeng Wang (Xi'an-Jiaotong Liverpool University)
TransformerLarge Language ModelTextMultimodality
🎯 What it does: The research proposes a model merging method called IP-Merging, which transfers mathematical reasoning capabilities from specialized mathematical LLMs to multimodal LLMs.
Can Multi-Modal LLMs Provide Live Step-by-Step Task Guidance?
Apratim Bhattacharyya (Qualcomm AI Research), Roland Memisevic (Qualcomm AI Research)
TransformerLarge Language ModelVision Language ModelVideoMultimodalityBenchmark
🎯 What it does: This paper presents a new dataset and benchmark—Qualcomm Interactive Cooking—for evaluating the capabilities of multimodal large language models in real-time, step-by-step guidance tasks, and trains and evaluates a new lightweight model, LIVEMAMBA, based on this benchmark;
Can NeRFs "See" without Cameras?
Chaitanya Amballa (University of Illinois Urbana-Champaign), Romit Roy Choudhury (University of Illinois Urbana-Champaign)
Neural Radiance FieldPoint Cloud
🎯 What it does: This study proposes EchoNeRF, which modifies the NeRF framework to utilize WiFi signal power (including LoS and first-order reflections) to infer indoor floor layouts;
Can We Infer Confidential Properties of Training Data from LLMs?
Pengrun Huang (University of California San Diego), Ruihan Wu (University of California San Diego)
Adversarial AttackTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringTextBenchmark
🎯 What it does: A benchmark task for property inference (PropInfer) targeting large language models is proposed, and attack methods are demonstrated under two fine-tuning modes (question-answering and chat completion).
Cancer Survival Analysis via Zero-shot Tumor Microenvironment Segmentation on Low-resolution Whole Slide Pathology Images
Jiao Tang (Nanjing University of Aeronautics and Astronautics), Daoqiang Zhang (Nanjing University of Aeronautics and Astronautics)
ClassificationSegmentationGraph Neural NetworkTransformerLarge Language ModelPrompt EngineeringImageBiomedical Data
🎯 What it does: An end-to-end cancer survival analysis framework called ZTSurv is proposed, which achieves zero-shot tumor microenvironment (TME) segmentation on low-resolution whole slide images (WSI) and constructs heterogeneous graphs to predict patient survival risk.
Caption This, Reason That: VLMs Caught in the Middle
Zihan Weng (McGill University), Pouya Bashivan (McGill University)
TransformerLarge Language ModelVision Language ModelMultimodalityBenchmarkChain-of-Thought
🎯 What it does: By constructing two benchmarks, PAM and CVR, this paper systematically evaluates the performance of current visual-language models in the three core cognitive abilities of perception, attention, and memory, and proposes methods to enhance spatial reasoning and multi-image reasoning through visual-text decoupling (self-generated descriptions) and fine-tuning with fine-grained LoRA.
Capturing Individual Human Preferences with Reward Features
Andre Barreto, Hugo Larochelle (Google DeepMind)
Recommendation SystemReinforcement Learning from Human FeedbackLarge Language ModelReinforcement LearningText
🎯 What it does: This paper proposes a reward model architecture (RFM) that quickly adapts to users by decomposing the reward function into shared features and user-specific linear weights. It learns reward features using pairwise preference data in an RLHF environment to complete the training and adaptation of personalized models.
Capturing Polysemanticity with PRISM: A Multi-Concept Feature Description Framework
Laura Kopf (Technische Universität Berlin), Oliver Eberle (Technische Universität Berlin)
TransformerLarge Language ModelText
🎯 What it does: The PRISM framework is proposed to generate multi-concept descriptions for the internal features of large language models (LLMs), addressing the limitation of single descriptions that cannot capture polysemanticity.
CAR-Flow: Condition-Aware Reparameterization Aligns Source and Target for Better Flow Matching
Chen Chen (Apple Inc), Alex Schwing (Apple Inc)
GenerationData SynthesisFlow-based ModelRectified FlowImageStochastic Differential EquationOrdinary Differential Equation
🎯 What it does: This paper proposes Condition-Aware Reparameterization for Flow Matching (CAR-Flow), which reduces the probability paths that the flow matching model needs to learn by applying lightweight condition-aware translations to the source and/or target distributions, thereby accelerating training and improving sample quality.
CARE: Decoding-Time Safety Alignment via Rollback and Introspection Intervention
Xiaomeng Hu (Alibaba Group), Tsung-Yi Ho (Chinese University of Hong Kong)
OptimizationSafty and PrivacyTransformerLarge Language ModelPrompt EngineeringContrastive LearningText
🎯 What it does: The CARE framework is proposed to achieve safety alignment during the LLM decoding process, primarily through two modules: the detect-rollback-intervention mechanism and self-reflection (Introspection) intervention.
CAS-Spec: Cascade Adaptive Self-Speculative Decoding for On-the-Fly Lossless Inference Acceleration of LLMs
Zhiyuan Ning (TeleAI), Xuelong Li (Shanghai Jiao Tong University)
Computational EfficiencyTransformerLarge Language ModelTextBenchmark
🎯 What it does: This paper proposes CAS-Spec, which constructs a multi-level self-inference model using Dynamic Switchable Inference Acceleration (DSIA) to achieve training-free LLM inference acceleration.
Cascaded Language Models for Cost-Effective Human–AI Decision-Making
Claudio Fanconi (University of Cambridge), Mihaela van der Schaar (University of Cambridge)
TransformerLarge Language ModelReinforcement LearningTextBiomedical Data
🎯 What it does: An adaptive decision-making framework based on multi-level LLMs and human experts is proposed, achieving model delegation, fallback, and online learning through confidence and uncertainty thresholds, balancing accuracy, cost, and fallback.
CAT: Circular-Convolutional Attention for Sub-Quadratic Transformers
Yoshihiro Yamada (Preferred Networks)
Computational EfficiencyTransformerImageText
🎯 What it does: Proposes the Circular-convolutional Attention (CAT) and Engineering-Isomorphic Transformers (EITs) framework, achieving O(N log N) global softmax-preserving attention;
CAT: Content-Adaptive Image Tokenization
Junhong Shen, Chunting Zhou
GenerationCompressionLarge Language ModelAuto EncoderGenerative Adversarial NetworkImage
🎯 What it does: A content-adaptive image tagger (CAT) is designed to dynamically determine the compression ratio based on the textual description of the image and uses a nested VAE to achieve multiple compression ratios with a single model.
CATransformers: Carbon Aware Transformers Through Joint Model-Hardware Optimization
Irene Wang (Georgia Institute of Technology), Bilge Acun (Meta)
OptimizationTransformerContrastive LearningMultimodality
🎯 What it does: A framework named CATransformers has been developed, capable of simultaneously optimizing the structure of Transformer models and the configuration of hardware accelerators during the early design phase to minimize total carbon emissions (including operational carbon and inherent carbon).
Causal Climate Emulation with Bayesian Filtering
Sebastian Hickman (University of Cambridge), Julien Boussard (McGill University)
Explainability and InterpretabilityRepresentation LearningAuto EncoderTime SeriesPhysics Related
🎯 What it does: A physical information and causally interpretable climate simulator has been developed, utilizing causal representation learning to extract potential causal structures, and achieving long-term stable climate projections and counterfactual experiments through Bayesian filtering.
Causal Differentiating Concepts: Interpreting LM Behavior via Causal Representation Learning
Navita Goyal (University of Maryland), Dhanya Sridhar (Mila-Quebec AI Institute)
Explainability and InterpretabilityRepresentation LearningTransformerAuto EncoderContrastive LearningText
🎯 What it does: This study investigates an unsupervised method to separate causally distinguishable concepts from language model activations to explain model behavior.
Causal Discovery and Inference through Next-Token Prediction
Eivinas Butkus (Columbia University), Nikolaus Kriegeskorte (Columbia University)
TransformerLarge Language ModelText
🎯 What it does: This paper demonstrates that a GPT-style Transformer trained solely on next-token prediction can learn causal structures from artificially generated linear Gaussian structural causal model (SCM) texts and can perform counterfactual reasoning on unseen SCMs.
Causal Discovery over Clusters of Variables in Markovian Systems
Tara Vafai Anand, Elias Bareinboim (Columbia University)
GraphTabular
🎯 What it does: This paper studies a new framework for causal discovery of predefined variable clusters in Markov systems, proposing two graphical models: α C-DAG and α C-CPDAG, and providing a complete learning algorithm called CLOC.
Causal Explanation-Guided Learning for Organ Allocation
Alessandro Marchese (Vrije Universiteit Brussel), Sam Verboven (Vrije Universiteit Brussel)
OptimizationExplainability and InterpretabilityAdversarial AttackContrastive LearningTabularBiomedical Data
🎯 What it does: This paper presents CLEXNET, a causal explanation-guided model that utilizes directional information about the reasons for rejection during organ transplantation to improve acceptance predictions.
Causal Head Gating: A Framework for Interpreting Roles of Attention Heads in Transformers
Andrew Joohun Nam, Sarah-Jane Leslie (Princeton University)
Explainability and InterpretabilityTransformerText
🎯 What it does: A scalable Causal Head Gate (CHG) method is proposed to identify the causal roles of different attention heads in Transformer models when completing tasks.
Causal LLM Routing: End-to-End Regret Minimization from Observational Data
Asterios Tsiourvas (Massachusetts Institute of Technology), Georgia Perakis (Massachusetts Institute of Technology)
Recommendation SystemOptimizationTransformerLarge Language ModelReinforcement LearningTextBenchmark
🎯 What it does: A Causal LLM routing framework that directly minimizes decision regret on observable data addresses the issues of traditional separation of prediction and routing, which requires complete feedback data.
Causal Mixture Models: Characterization and Discovery
Sarah Mameche (CISPA Helmholtz Center for Information Security), Jilles Vreeken (CISPA Helmholtz Center for Information Security)
Mixture of ExpertsTabular
🎯 What it does: A causal mixture model based on conditional linear mixed regression is proposed and embedded in score-based causal structure search.
Causal Spatio-Temporal Prediction: An Effective and Efficient Multi-Modal Approach
Yuting Huang (Zhejiang University), Yunjun Gao (Zhejiang University)
OptimizationComputational EfficiencyGraph Neural NetworkContrastive LearningMultimodalityTime Series
🎯 What it does: Proposed the E^2-CSTP framework to achieve multimodal spatiotemporal prediction.
Causal Sufficiency and Necessity Improves Chain-of-Thought Reasoning
Xiangning Yu (Tianjin University), Mengyue Yang (Hong Kong University of Science and Technology)
TransformerLarge Language ModelSupervised Fine-TuningTextChain-of-Thought
🎯 What it does: Proposes a chain-of-thought (CoT) optimization framework based on probabilistic necessary and sufficient (PNS) reasoning, which automatically eliminates redundant steps while maintaining or improving accuracy.
Causal-R: A Causal-Reasoning Geometry Problem Solver for Optimized Solution Exploration
Wenjun Wu (Xi'an Jiaotong University), Jun Liu (Lenovo Research)
OptimizationExplainability and InterpretabilityComputational EfficiencyGraph Neural NetworkGraphBenchmark
🎯 What it does: The Causal-R model is proposed, which achieves efficient problem solving and multi-solution exploration for geometric problems through causal graph reasoning and forward matrix reasoning.
Causality Meets Locality: Provably Generalizable and Scalable Policy Learning for Networked Systems
Hao Liang (King's College London), Yali Du (King's College London)
Meta LearningReinforcement LearningTabularBenchmark
🎯 What it does: In large-scale networked systems, the GSAC framework is proposed, combining causal representation learning with meta actor-critic to achieve scalable and generalizable multi-agent reinforcement learning.
Causality Meets the Table: Debiasing LLMs for Faithful TableQA via Front-Door Intervention
Zhen Yang (Anhui University), Shu Zhao (Anhui University)
TransformerLarge Language ModelAgentic AITabular
🎯 What it does: Proposes the CIT framework, which eliminates confounding bias caused by vocabulary co-occurrence in TableQA through front-door adjustment;
Causality-Induced Positional Encoding for Transformer-Based Representation Learning of Non-Sequential Features
Kaichen Xu (Emory University), Xiaobo Sun (Emory University)
Representation LearningTransformerContrastive LearningTabularBiomedical Data
🎯 What it does: Proposes the CAPE method, which provides location-aware encoding for Transformer to handle non-sequential but causally related features;
Causally Reliable Concept Bottleneck Models
Giovanni De Felice (Università della Svizzera Italiana), Alberto Termine (Scuola Universitaria Professionale della Svizzera Italiana)
Large Language ModelImageTabularRetrieval-Augmented Generation
🎯 What it does: This paper proposes Causally reliable Concept Bottleneck Models (C2BMs), which structure the concept bottleneck into a structured causal model (SCM) that aligns with true causal mechanisms, and provides an automated process for concept discovery and causal graph construction.
CausalPFN: Amortized Causal Effect Estimation via In-Context Learning
Vahid Balazadeh (University of Toronto), Rahul Krishnan
TransformerTabular
🎯 What it does: CausalPFN is proposed, a transformer-based pre-trained model that can estimate causal effects in one go.
CausalVTG: Towards Robust Video Temporal Grounding via Causal Inference
Qiyi Wang, Ying Shen
RetrievalContrastive LearningVideo
🎯 What it does: This paper proposes the CausalVTG framework, which utilizes causal reasoning to eliminate co-occurrence bias in video temporal localization and introduces counterfactual inference to determine whether a query is truly related to the video.
CCL: Causal-aware In-context Learning for Out-of-Distribution Generalization
Hoyoon Byun (Yonsei University), Kyungwoo Song (Yonsei University)
Domain AdaptationRepresentation LearningTransformerLarge Language ModelAuto EncoderText
🎯 What it does: A causal-aware context learning (CCL) framework is proposed for out-of-distribution (OOD) scenarios, utilizing causal representations to select examples and enhance the generalization performance of large language models.
CCS: Controllable and Constrained Sampling with Diffusion Models via Initial Noise Perturbation
Bowen Song (University of Michigan), Liyue Shen (Rutgers University)
GenerationData SynthesisDiffusion modelImageOrdinary Differential Equation
🎯 What it does: This study investigates the linear impact of initial noise perturbations in diffusion models on the generated results and proposes a Controlled and Constrained Sampling (CCS) method to achieve precise sampling under given target mean and mean squared error (MSE).
CDFlow: Building Invertible Layers with Circulant and Diagonal Matrices
XUCHEN FENG, Siyu Liao (Sun Yat-sen University)
GenerationComputational EfficiencyFlow-based ModelImage
🎯 What it does: CDFlow is proposed, a regularized flow generative model based on the alternating multiplication of circulant and diagonal matrices in an invertible linear layer;
CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning
MingYu Lu, Su-In Lee (University of Washington)
RetrievalDrug DiscoveryTransformerContrastive LearningImageTextMultimodalityBiomedical Data
🎯 What it does: We propose CellCLIP, a cross-modal contrastive learning framework that uses natural language to encode perturbations and aligns perturbations with morphology by combining them with Cell Painting images, improving cross-modal retrieval and biological downstream tasks.
Centralized Reward Agent for Knowledge Sharing and Transfer in Multi-Task Reinforcement Learning
Haozhe Ma (National University of Singapore), Tze-Yun Leong (Nanyang Technological University)
Reinforcement Learning
🎯 What it does: The CenRA framework is proposed, which combines reward shaping with multi-task reinforcement learning. A centralized reward agent (CRA) distributes dense rewards to enhance learning efficiency in sparse reward environments and support new task transfer.
Certifying Concavity and Monotonicity in Games via Sum-of-Squares Hierarchies
Vincent Leon (University of Illinois Urbana-Champaign), Antonios Varvitsiotis (Singapore University of Technology and Design)
Optimization
🎯 What it does: The study verifies the concavity and monotonicity of multi-player games by constructing and applying sum-of-squares hierarchies, and provides optimal approximations of SOS concave/monotonic games.
Certifying Deep Network Risks and Individual Predictions with PAC-Bayes Loss via Localized Priors
Wen Dong (Air Force Research Laboratory)
ClassificationOptimizationExplainability and InterpretabilityConvolutional Neural NetworkImageText
🎯 What it does: This paper proposes a certifiable approach for risk and individual prediction in deep networks based on localized PAC-Bayes priors.
Certifying Stability of Reinforcement Learning Policies using Generalized Lyapunov Functions
Kehan Long (University of California San Diego), Nikolay Atanasov (University of California San Diego)
Reinforcement Learning
🎯 What it does: A method has been designed and implemented that utilizes reinforcement learning value functions and neural residuals to construct generalized Lyapunov functions, which can perform post-hoc stability verification of trained RL policies and support joint learning of controllers and certificates.
CF-VLM：CounterFactual Vision-Language Fine-tuning
Jusheng Zhang (Sun Yat-sen University), Keze Wang (Sun Yat-sen University)
ClassificationRecognitionRetrievalTransformerVision Language ModelContrastive LearningImageTextMultimodalityChain-of-Thought
🎯 What it does: A CF-VLM framework is designed to enhance the causal reasoning and fine-grained discrimination of visual-text models by generating adversarial samples with minimal causal interventions on the original image text, providing three complementary training objectives.
CG-SSL: Concept-Guided Self-Supervised Learning
Sara Atito (Surrey Institute for People-Centred AI), Muhammad Awais (Surrey Institute for People-Centred AI)
SegmentationRepresentation LearningTransformerContrastive LearningImage
🎯 What it does: A three-stage self-supervised learning framework is proposed: first, global representations are learned through iBoT; then, a cross-attention clustering module is used to discover visual concept Tokens; finally, geometric projection is employed to align concepts from different views, achieving interpretable and spatially consistent features.
CGS-GAN: 3D Consistent Gaussian Splatting GANs for High Resolution Human Head Synthesis
Florian Barthel (Fraunhofer Heinrich Hertz Institute), Peter Eisert (Fraunhofer Heinrich Hertz Institute)
GenerationData SynthesisTransformerGenerative Adversarial NetworkGaussian SplattingImage
🎯 What it does: Designed and trained CGS-GAN, utilizing 3D Gaussian splatting to achieve high-resolution, 3D consistent face head synthesis;
ChA-MAEViT: Unifying Channel-Aware Masked Autoencoders and Multi-Channel Vision Transformers for Improved Cross-Channel Learning
Chau Pham (Boston University), Bryan A. Plummer (Boston University)
ClassificationRepresentation LearningTransformerAuto EncoderImage
🎯 What it does: This study proposes a Vision Transformer model based on Masked Autoencoder, named ChA-MAEViT, for multi-channel images (MCI), aiming to enhance cross-channel feature learning and self-supervised representation.
Chain of Execution Supervision Promotes General Reasoning in Large Language Models
Nuo Chen (Alibaba), Dayiheng Liu (Alibaba)
AI Code AssistantTransformerLarge Language ModelSupervised Fine-TuningPrompt EngineeringText
🎯 What it does: A TracePile dataset was constructed, utilizing the code execution process to generate fine-grained Chain of Execution (CoE) explanations to enhance the general reasoning capabilities of large language models.
Chain-of-Action: Trajectory Autoregressive Modeling for Robotic Manipulation
Wenbo Zhang (University of Adelaide), Xiao Ma (ByteDance)
Robotic IntelligenceTransformerReinforcement LearningSequential
🎯 What it does: A visual motion strategy based on reverse trajectory autoregression, called Chain-of-Action (CoA), is proposed, which achieves global-to-local reasoning by generating a complete action sequence in reverse from task keyframes.
Chain-of-Model Learning for Language Model
Xiaohua Wang (Microsoft Research), Lili Qiu (Fudan University)
TransformerLarge Language ModelText
🎯 What it does: The paper proposes a Chain-of-Model learning framework that embeds multi-scale information into the hidden layers of a Transformer, achieving a scalable multi-scale language model CoLM and CoLM-Air, which supports flexible inference, pre-fill acceleration, and continuous training.
Chain-of-Retrieval Augmented Generation
Liang Wang (Microsoft Research), Furu Wei (Microsoft Research)
GenerationRetrievalTransformerLarge Language ModelTextRetrieval-Augmented Generation
🎯 What it does: A framework named CoRAG has been designed and implemented, allowing large language models to retrieve information step-by-step and dynamically rewrite queries to form a retrieval chain before generating answers.
Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment
Bryan Sangwoo Kim (Korea Advanced Institute of Science and Technology), Jong Chul Ye (Korea Advanced Institute of Science and Technology)
Super ResolutionReinforcement LearningVision Language ModelDiffusion modelImage
🎯 What it does: The Chain-of-Zoom (CoZ) framework is proposed, which decomposes the extreme magnification task of a single image into a series of autoregressive inferences at intermediate scales, enabling the extension of existing SR models to extreme magnifications such as 64× and 256× without retraining;